Build-your-own NAT KVM servers from $0.000135 an hour, with a private /24, two forwarded TCP ports and a shared HAProxy — cheap, hourly, and explicitly still in Alpha.
Short answer: One always-on personal agent (2 vCPU / 4 GiB / 20 GiB) costs about $2.98 a month here, the cheapest of the 24 providers that run it as described. The AI model adds $7.30 to $488.07 a month on top, depending on the model: MiniMax M3.1 Flash at the low end, $204.19 or more on a frontier model such as Claude Opus 5.5, up to GPT-6 Astra. That assumes 144 million tokens a month (30 requests a day × 160k tokens × 30 days, our assumption) at the median zero-retention price of each model. Fits well: A throwaway box for an experiment, a test deploy or a CI job: $0.003886 an hour for 4 GiB and 10 GB, deleted when the job ends.
tierhive.com29 providersSources linked
TierHive rents KVM virtual machines by the hour at the lowest end of the market. You do not pick from a plan list: you move sliders for RAM (128 MB to 8 GB in 128 MB steps), NVMe (1 to 50 GB) and, in four locations, HDD up to 4 TB, then choose a CPU-priority, IOPS and network tier. The smallest configuration, 128 MB of RAM with 1 GB of disk, is $0.000135 an hour — about ten cents for a 720-hour month.
The unusual part is the network. A server does not get its own public IPv4. Each account gets a private /24 that is meshed across locations, and each server gets two forwarded TCP ports and two UDP ports on the public address of the machine it runs on. The first TCP port is always mapped to the server's port 22, so SSH arrives on a random high port of a shared IP. Web traffic normally goes through the free shared HAProxy instead: you add a domain in the panel, point DNS at TierHive, and it terminates TLS with a Let's Encrypt certificate. An IPv6 /64 per location is also available, up to ten addresses per server, which opens nearly all ports and removes the NAT limit.
vCPU is not something you buy. A server with 2 GB of RAM or less gets one vCPU and anything larger gets two, so two is the ceiling. The CPU is shared and weighted by the priority tier you pay for, and a server that averages a full core for two hours is cut to half a core until it reboots.
Billing is prepaid: one token is one US dollar, the first top-up is $3, and each server bills its hourly rate with a one-hour minimum. The operator says a server that is powered off bills only its storage, and that deleting it ends the bill immediately. There is no monthly price and no monthly cap.
TierHive is candid about what it is not. Its acceptable-use policy says the service is in Alpha, that data loss can occur at any time without warning, that backups are the customer's responsibility, and that there is no SLA. The About page says it is a budget service on older hardware, that it is not the place to host anything you need uptime for, and that if data is lost you get a fresh, blank server rather than a recovery.
At a glance
Who it is for
Readers who want the cheapest possible Linux box for a hobby agent, a bot, a test box or a throwaway node, who can work a console, and who keep their own backups. Not for production, not for anything that needs a public IPv4 of its own, more than two vCPUs, or a promise about uptime.
Isolation
A KVM virtual machine with its own kernel and root, on the operator's own QEMU/KVM/libvirt stack. CPU and disk I/O are shared and weighted, not dedicated; a server that averages a full core for two hours is throttled to half a core until it is rebooted.
Sizes
Build your own: RAM 128 MB to 8 GB in 128 MB steps, NVMe 1 to 50 GB, HDD up to 4,000 GB in London, Singapore, Sydney and Vint Hill, plus backup and snapshot storage. vCPU is not a choice — one core up to 2 GB of RAM, two cores above it, never more. 8 GB is not available in every location (Tokyo tops out at 1 GB). 19 locations across Europe, North America, South America, Asia-Pacific and Africa.
A personal agent here
About $2.98 a month for one always-on 2 vCPU / 4 GiB machine (4 GiB / 20 GB: 2 vCPU / 4 GiB / 20 GiB), before the AI model. Change the numbers
Our reading of the sourced material on this page, not a provider claim.
Fits well
A throwaway box for an experiment, a test deploy or a CI job: $0.003886 an hour for 4 GiB and 10 GB, deleted when the job ends.
A hobby agent, Discord or Slack bot, or small MCP server that needs a cheap always-on Linux box and can live behind a forwarded port or the shared HAProxy.
Three hours a day of work on a machine that keeps its files: about $0.62 a month at 4 GiB and 20 GB, most of it storage while stopped.
A homelab spread across 19 locations with a private /24 meshed between them, for WireGuard relays, monitoring probes or redundant small services.
Fits badly
Anything that needs its own public IPv4, more than two forwarded TCP ports, or a fixed address other services can allowlist.
Anything with four or more vCPUs, more than 8 GB of RAM, a GPU or a dedicated core.
Production, or data you cannot lose: the operator's own policy says data loss can occur at any time and that there is no SLA.
A team that needs sub-users, SSO or an audit trail, or a buyer who needs a SOC 2 report or a data processing agreement.
Scraping at scale: the acceptable-use policy forbids automated scraping of external sites without prior written approval.
How it bills
Model
Prepaid hourly tokens at $1 each. No subscription and no monthly plan: you add credit and each server draws its hourly rate until you delete it. The first top-up is $3.
Meter
A per-hour price built from published resource rates — RAM $0.0009216 per GiB-hour, NVMe $0.00002 per GB-hour, HDD $0.000007, backup storage $0.000005, snapshots $0.00002 — plus optional priority tiers at $0.00002 (Medium) or $0.00003 (High) an hour each. The panel's own estimate uses 720 hours for a month.
While stopped
The operator says a powered-off server bills only its storage, and that deleting it stops the charge. A stop is a power-off, so memory and running processes are lost; nothing resumes where it left off. The stopped-state details are only in the operator's forum offers, not on tierhive.com.
Plan fee
There is no platform fee and no plan. The only extra charges in the price feed are the DHCP fee (about $0.0007 a month), backup creation and restore fees, and backup or snapshot storage.
Minimum
One hour per server, and a $3 first top-up. A ten-minute server costs an hour; 1,000 ten-minute servers cost 1,000 hours.
The cheapest hourly Linux box in this guide: $0.000135 an hour for 128 MB and 1 GB, $0.003886 for 4 GiB and 10 GB.
Genuine per-resource pricing — RAM and disk are sliders, not plan steps — so a small machine is not forced to buy a large disk.
Storage-only billing while a server is powered off, which makes a three-hours-a-day machine cost cents rather than dollars.
A private /24 meshed across 19 locations, free shared HAProxy with Let's Encrypt, IPv6 /64s, and one-click recipes including Docker, WireGuard and WordPress.
An unusually candid operator: the limits, the lack of an SLA and the risk of data loss are all written down in public.
Using it without writing code
Dashboard
A custom control panel behind a login: create and rebuild servers, change RAM and disk, start and stop, map ports on the Subnet Config page, allocate IPv6, schedule backups and snapshots, add domains to HAProxy, and open a noVNC console.
Without code
No. Creating a server is click-through and there are one-click recipes for Docker, WordPress, WireGuard and similar, but the recipes end at a console, ports are mapped by hand, and the HAProxy backend is your own private IP. Nothing here sets up an AI agent for you.
Signup
A username, email and password, then a six-digit code at every sign-in. No card is needed to register; the first tokens are bought through FastSpring with a $3 minimum. The terms forbid disposable email addresses and VPN sign-up, although the About page says the opposite.
Docs and support
Thin and blog-shaped: an About page, the acceptable-use policy, and 18 blog posts of which about a dozen are how-tos (finding your ports, HAProxy, IPv6, growing a disk, chrooting a backup). There is no knowledge base, no FAQ and no API reference.
Tools you can use
TierHive has 3 of the 10 ways to drive a machine that this guide checks. From its own documentation; how pleasant each one is to use was not tested. Compare every provider
Any size you likeWith a catch RAM and disk are free-form sliders, but CPU is not a choice — one core up to 2 GB and two above that — so the machine is only half build-your-own.
Resize while runningWith a catch The panel changes RAM and disk while the server runs, but the guest only sees a bigger disk after a reboot, and vCPU cannot be changed at all.
Reusable templatesWith a catch You can rebuild from the OS list or run a community recipe, but a configured server cannot be saved as your own template.
Non-web portsWith a catch Each server gets two forwarded TCP ports and SSH takes one, so only one TCP service is reachable from outside unless you use IPv6.
Fixed outgoing IPWith a catch Outbound IPv4 leaves from the host's address, shared with every other server on that machine; only the IPv6 addresses are yours.
Hard spending capWith a catch Spending is limited by the prepaid token balance rather than a limit you set, and nothing says what happens to a running server when the balance reaches zero.
None. The acceptable-use policy says there is no SLA and that data loss can occur at any time.
“We do not guarantee any specific uptime or Service Level Agreement (SLA)”tierhive.comdocs-confirmed
Status page
Per-location 90-day uptime, reading 99.99% in Amsterdam, London and Singapore, 99.91% in Tokyo and 100% in the other fifteen locations on 2026-11-04.
“Amsterdam, NL Operational 90 days ago 99.99 % uptime Today”tierhive.comdocs-confirmed
Who takes the payment
FastSpring is the merchant of record, so FastSpring is the seller on the invoice rather than Kernel Keepers.
“All payments are processed securely through our Merchant of Record, FastSpring.”tierhive.comdocs-confirmed
Caveats and contradictions
Shown as the research found them. A contradiction means the provider's own pages disagree; an open question means the answer is not published.
01
Alpha, with data loss written into the policy Open question
The acceptable-use policy states that TierHive is in Alpha, that data loss can occur at any time without warning, that backups are the customer's sole responsibility and that those backups may also be lost. It tells customers not to use the service for production or critical data. The About page adds that if data is lost you get a fresh, blank server, not a recovery.
Storage-only billing while stopped is not on the provider's website Open question
The operator's forum offers say a powered-off server bills only storage, and the pricing feed has no separate stopped rate. Nothing on tierhive.com describes the stopped state, so the card bills the NVMe rate for stopped hours and treats the rest as zero. If stopped hours billed at the running rate, C would rise from $0.62 to $2.98 (730 h × $0.0040864).
“You only pay for storage when the VPS is powered off.”
The site calls the largest machine 8 GB while the price feed calls the same ceiling 8,192 MB, so its MB and GB are binary and its per-GB-hour disk rate maps 1:1 to this site's GiB. For RAM that is direct; for disk it is an inference, because no page defines the unit in bytes. If the disk GB is decimal, every disk charge is 7.37% higher — A $3.90, B $6.17, C $0.65 — and a 50 GiB disk would sit above the 50 GB ceiling.
“Up to 8GB RAM per VPS (not all locations support 8GB)”
Two forwarded TCP ports, and SSH takes one of them Open question
Every server gets two TCP and two UDP ports on the public address of its host, and the first TCP port is always mapped to port 22. That leaves one spare TCP port for anything else, which is why the provider's own HAProxy how-to uses a second VPS as a bastion. An IPv6 /64 per location, up to ten addresses per server, is the way around it — but the shared HAProxy does not serve IPv6.
“Your nat VPS is provided with 2 open TCP ports and 2 UDP port. By design the 1st TCP port is mapped to your VPS port 22/TCP.”
Free trial and a $3 minimum top-up are both advertised Contradiction
The homepage says a $3 minimum top-up is required after registration to use the service; the blog says to start with the free trial and no credit card; the price feed grants $0.01 of sign-up credit; the forum offers promise free tokens to try before buying. The honest reading is a free account with one cent of credit that may not be usable until $3 is added.
“A $3 minimum top up is required after registration to use this service.”
The terms ban what the About page invites Contradiction
The terms forbid disposable email addresses and signing up over a VPN or proxy; the About page says the operator has no objection to a free or disposable email or a VPN. The terms also describe monthly or annual subscriptions, which is boilerplate from the company's other products — the actual product is prepaid hourly tokens.
“You may not use free or disposable email addresses to create or maintain an Account.”
A Sao Paulo location sits next to a ban on Brazilian customers Contradiction
The price feed lists Sao Paulo with a 4,096 MB ceiling, while the terms say the service cannot be provided to customers in Brazil (along with China, Russia, Vietnam and Iran). The node is real; customers located in Brazil are refused.
The marketing blog quotes prices that no longer exist Open question
The hourly-VPS explainer quotes a Popular tier at $0.004061/h, which reproduces from no current size, while the homepage's Popular card is $0.007561/h. Its Standard figure, $0.002343/h, still matches. No rate in this package comes from the blog.
“The Popular tier is $0.004061/hour. The Standard is $0.002343/hour.”
The site's sitemap lists every public page and there is no developer, API or CLI page; every guessed panel path returns 401 behind the login. The only public JSON endpoints are the marketing calculator's rate feed and the recipe ticker. A third-party dossier reaches the same conclusion. Nothing here says an API will never exist — only that none is documented today.
10
Shared CPU with a watchdog and a fair-use network throttle Open question
The provider warns that CPU is the most contended resource, that steal will be higher than on an enterprise host, and that a server averaging a full core for two hours is limited to half a core until it reboots. Network speed is a tier (50 to 300 Mbit/s by location) that steps down by 10 Mbit/s an hour under sustained use, to a floor of 10 Mbit/s. Neither binds at the loads in A, B and C.
“We have a CPU watchdog: if you average 100% of a CPU core for more than 2 hours solid, you will be limited to 50% of a core on the backend.”
No SOC 2 report or data processing agreement is published Open question
The legal pages are terms, privacy, refunds and acceptable use. There is no DPA, no sub-processor list and no audit report. The company is UK-based; its EU locations are in Germany, France, the Netherlands and Poland. A buyer who needs GDPR paperwork should ask before buying.
List-price illustrations from the calculator, not quotes. Prices checked 4 November 2026.
Billed in whole hoursEach server costs at least an hour, so a ten-minute task pays for sixty.
Shared CPU, 2 vCPU at mostCPU is weighted between customers rather than dedicated, and a server that averages a full core for two hours is cut to half a core until it is rebooted.
A · 1,000 short tasks
$3.89
4 GiB / 10 GB: 2 vCPU / 4 GiB / 10 GiB (plan: Pay as you go)
Machine time (4 GiB / 10 GB)1,000 runs × 1 h each, rounded up × $0.003886$3.89
Prepaid tokens at $1 each, $3 minimum top-up. No subscription, no monthly cap: a server left running bills its hourly rate for as long as it exists.
Rates come from the calculator's own data feed, https://tierhive.com/api/pricing.php: RAM 0.00000090 tokens per MB-hour, NVMe 0.00002000 per GB-hour, HDD 0.00000700 per GB-hour, backup storage 0.00000500 per GB-hour, snapshots 0.00002000 per GB-hour, and a DHCP charge of 0.00000100 per hour plus a one-off activation. One token is $1.
Each shape's hourly price is those rates applied to its RAM and NVMe with low-priority CPU, IOPS and network, and reproduces the provider's own cards (128 MB / 1 GB = 0.000135, 2 GB / 25 GB = 0.002343, 4 GB / 50 GB = 0.004686). The card lists seven representative points because TierHive is build-your-own: RAM is any 128 MB step from 128 MB to 8 GB, NVMe any GB from 1 to 50. Priority tiers add 0.00002/h each for Medium and 0.00003/h for High, which is not included here.
Units: TierHive's calculator runs its RAM slider in MB and the site calls 8 GB what the feed calls 8,192 MB, so its MB and GB are binary and a per-GB-hour rate is treated here as per GiB-hour, with each shape's disk size mapped 1:1 to the engine's GiB. That mapping is an inference for disk, since no page defines the unit in bytes. If the disk GB is decimal, every disk charge rises by 7.37% (A $3.90, B $6.17, C $0.65); disk.decimalGB is deliberately not set, because the provider's own 8 GB = 8,192 MB labelling points to binary units.
vCPU is not a control: it is 1 up to 2 GiB of RAM and 2 above that, so a 4 vCPU machine cannot be bought at any price. Workload B is therefore priced on the 8 GiB shape the provider would sell, and the engine flags the shortfall.
A one-hour minimum per server is modelled with wholeHours. The operator states it in its offer post: "You can create a VPS or 253 VPS with a 1-hour minimum and you are only billed for that hour". Whether a 90-minute server bills 1.5 or 2 hours is not published; A, B and C are unaffected.
While powered off the card bills disk only, at the same NVMe rate as when running. That follows the operator's offer posts ("You only pay for storage when the VPS is powered off") rather than any page on tierhive.com, and the exact stopped-state charges — backup storage, the DHCP fee, whether RAM is really dropped — are not published. If stopped hours were billed at the running rate, C would cost $2.98 instead of $0.62; B is unaffected.
Nothing caps a month: there is no monthly price, so an always-on server costs 730 hours of its hourly rate. The provider's own monthly estimates use 720 hours (24 × 30), which is why its cards read a little lower than 730 hours here.
Not priced: the DHCP charge (about $0.0007 a month), backup creation and restore fees ($0.00002 each), backup and snapshot storage, HDD at 0.000007/GB-h in London, Singapore, Sydney and Vint Hill, and a network-backed storage rate in the feed with no documented product.
The price feed lists 19 locations with per-location RAM ceilings (8 GB in Amsterdam, Beauharnois, Frankfurt, Gravelines, London, New York, Sydney and Vancouver; 1 GB in Tokyo) and speed limits from 50 to 300 Mbit/s. Only one region is modelled here, since price does not vary by location.
This is an Alpha service: the operator's own policy says data loss can occur without warning and that there is no SLA. The price is real, the durability is not promised.
B · One always-on box
$6.11, with a catch
8 GiB / 50 GB: 2 vCPU / 8 GiB / 50 GiB (plan: Pay as you go)
Machine time (8 GiB / 50 GB)730 h × $0.008373$6.11
Prepaid tokens at $1 each, $3 minimum top-up. No subscription, no monthly cap: a server left running bills its hourly rate for as long as it exists.
Rates come from the calculator's own data feed, https://tierhive.com/api/pricing.php: RAM 0.00000090 tokens per MB-hour, NVMe 0.00002000 per GB-hour, HDD 0.00000700 per GB-hour, backup storage 0.00000500 per GB-hour, snapshots 0.00002000 per GB-hour, and a DHCP charge of 0.00000100 per hour plus a one-off activation. One token is $1.
Each shape's hourly price is those rates applied to its RAM and NVMe with low-priority CPU, IOPS and network, and reproduces the provider's own cards (128 MB / 1 GB = 0.000135, 2 GB / 25 GB = 0.002343, 4 GB / 50 GB = 0.004686). The card lists seven representative points because TierHive is build-your-own: RAM is any 128 MB step from 128 MB to 8 GB, NVMe any GB from 1 to 50. Priority tiers add 0.00002/h each for Medium and 0.00003/h for High, which is not included here.
Units: TierHive's calculator runs its RAM slider in MB and the site calls 8 GB what the feed calls 8,192 MB, so its MB and GB are binary and a per-GB-hour rate is treated here as per GiB-hour, with each shape's disk size mapped 1:1 to the engine's GiB. That mapping is an inference for disk, since no page defines the unit in bytes. If the disk GB is decimal, every disk charge rises by 7.37% (A $3.90, B $6.17, C $0.65); disk.decimalGB is deliberately not set, because the provider's own 8 GB = 8,192 MB labelling points to binary units.
vCPU is not a control: it is 1 up to 2 GiB of RAM and 2 above that, so a 4 vCPU machine cannot be bought at any price. Workload B is therefore priced on the 8 GiB shape the provider would sell, and the engine flags the shortfall.
A one-hour minimum per server is modelled with wholeHours. The operator states it in its offer post: "You can create a VPS or 253 VPS with a 1-hour minimum and you are only billed for that hour". Whether a 90-minute server bills 1.5 or 2 hours is not published; A, B and C are unaffected.
While powered off the card bills disk only, at the same NVMe rate as when running. That follows the operator's offer posts ("You only pay for storage when the VPS is powered off") rather than any page on tierhive.com, and the exact stopped-state charges — backup storage, the DHCP fee, whether RAM is really dropped — are not published. If stopped hours were billed at the running rate, C would cost $2.98 instead of $0.62; B is unaffected.
Nothing caps a month: there is no monthly price, so an always-on server costs 730 hours of its hourly rate. The provider's own monthly estimates use 720 hours (24 × 30), which is why its cards read a little lower than 730 hours here.
Not priced: the DHCP charge (about $0.0007 a month), backup creation and restore fees ($0.00002 each), backup and snapshot storage, HDD at 0.000007/GB-h in London, Singapore, Sydney and Vint Hill, and a network-backed storage rate in the feed with no documented product.
The price feed lists 19 locations with per-location RAM ceilings (8 GB in Amsterdam, Beauharnois, Frankfurt, Gravelines, London, New York, Sydney and Vancouver; 1 GB in Tokyo) and speed limits from 50 to 300 Mbit/s. Only one region is modelled here, since price does not vary by location.
This is an Alpha service: the operator's own policy says data loss can occur without warning and that there is no SLA. The price is real, the durability is not promised.
It cannot run this exactly as described: No size has 4 vCPU and 8 GiB; the largest is 2 vCPU / 8 GiB.
C · Three hours a day
$0.62
4 GiB / 20 GB: 2 vCPU / 4 GiB / 20 GiB (plan: Pay as you go)
Machine time (4 GiB / 20 GB)90 h × $0.004086$0.37
Disk while stopped640 h × 20 GiB × $0.00002$0.26
Prepaid tokens at $1 each, $3 minimum top-up. No subscription, no monthly cap: a server left running bills its hourly rate for as long as it exists.
Rates come from the calculator's own data feed, https://tierhive.com/api/pricing.php: RAM 0.00000090 tokens per MB-hour, NVMe 0.00002000 per GB-hour, HDD 0.00000700 per GB-hour, backup storage 0.00000500 per GB-hour, snapshots 0.00002000 per GB-hour, and a DHCP charge of 0.00000100 per hour plus a one-off activation. One token is $1.
Each shape's hourly price is those rates applied to its RAM and NVMe with low-priority CPU, IOPS and network, and reproduces the provider's own cards (128 MB / 1 GB = 0.000135, 2 GB / 25 GB = 0.002343, 4 GB / 50 GB = 0.004686). The card lists seven representative points because TierHive is build-your-own: RAM is any 128 MB step from 128 MB to 8 GB, NVMe any GB from 1 to 50. Priority tiers add 0.00002/h each for Medium and 0.00003/h for High, which is not included here.
Units: TierHive's calculator runs its RAM slider in MB and the site calls 8 GB what the feed calls 8,192 MB, so its MB and GB are binary and a per-GB-hour rate is treated here as per GiB-hour, with each shape's disk size mapped 1:1 to the engine's GiB. That mapping is an inference for disk, since no page defines the unit in bytes. If the disk GB is decimal, every disk charge rises by 7.37% (A $3.90, B $6.17, C $0.65); disk.decimalGB is deliberately not set, because the provider's own 8 GB = 8,192 MB labelling points to binary units.
vCPU is not a control: it is 1 up to 2 GiB of RAM and 2 above that, so a 4 vCPU machine cannot be bought at any price. Workload B is therefore priced on the 8 GiB shape the provider would sell, and the engine flags the shortfall.
A one-hour minimum per server is modelled with wholeHours. The operator states it in its offer post: "You can create a VPS or 253 VPS with a 1-hour minimum and you are only billed for that hour". Whether a 90-minute server bills 1.5 or 2 hours is not published; A, B and C are unaffected.
While powered off the card bills disk only, at the same NVMe rate as when running. That follows the operator's offer posts ("You only pay for storage when the VPS is powered off") rather than any page on tierhive.com, and the exact stopped-state charges — backup storage, the DHCP fee, whether RAM is really dropped — are not published. If stopped hours were billed at the running rate, C would cost $2.98 instead of $0.62; B is unaffected.
Nothing caps a month: there is no monthly price, so an always-on server costs 730 hours of its hourly rate. The provider's own monthly estimates use 720 hours (24 × 30), which is why its cards read a little lower than 730 hours here.
Not priced: the DHCP charge (about $0.0007 a month), backup creation and restore fees ($0.00002 each), backup and snapshot storage, HDD at 0.000007/GB-h in London, Singapore, Sydney and Vint Hill, and a network-backed storage rate in the feed with no documented product.
The price feed lists 19 locations with per-location RAM ceilings (8 GB in Amsterdam, Beauharnois, Frankfurt, Gravelines, London, New York, Sydney and Vancouver; 1 GB in Tokyo) and speed limits from 50 to 300 Mbit/s. Only one region is modelled here, since price does not vary by location.
This is an Alpha service: the operator's own policy says data loss can occur without warning and that there is no SLA. The price is real, the durability is not promised.